Towards a Model and Specification for Visual Programming of

Sensors & Transducers Volume 4, Special Issue March 2009

www.sensorsportal.com

ISSN 1726-5479

Guest Editors: Dr. Pierre Payeur and Dr. Emil M. Petriu, University of Ottawa, Ottawa, ON, Canada Editor-in-Chief: professor Sergey Y. Yurish, phone: +34 696067716, fax: +34 93 4011989, e-mail: [email protected] Editors for Western Europe Meijer, Gerard C.M., Delft University of Technology, The Netherlands Ferrari, Vittorio, Universitá di Brescia, Italy Editor South America Costa-Felix, Rodrigo, Inmetro, Brazil

Editors for North America Datskos, Panos G., Oak Ridge National Laboratory, USA Fabien, J. Josse, Marquette University, USA Katz, Evgeny, Clarkson University, USA Editor for Asia Ohyama, Shinji, Tokyo Institute of Technology, Japan

Editor for Eastern Europe Sachenko, Anatoly, Ternopil State Economic University, Ukraine

Editor for Asia-Pacific Mukhopadhyay, Subhas, Massey University, New Zealand

Editorial Advisory Board Abdul Rahim, Ruzairi, Universiti Teknologi, Malaysia Ahmad, Mohd Noor, Nothern University of Engineering, Malaysia Annamalai, Karthigeyan, National Institute of Advanced Industrial Science and Technology, Japan Arcega, Francisco, University of Zaragoza, Spain Arguel, Philippe, CNRS, France Ahn, Jae-Pyoung, Korea Institute of Science and Technology, Korea Arndt, Michael, Robert Bosch GmbH, Germany Ascoli, Giorgio, George Mason University, USA Atalay, Selcuk, Inonu University, Turkey Atghiaee, Ahmad, University of Tehran, Iran Augutis, Vygantas, Kaunas University of Technology, Lithuania Avachit, Patil Lalchand, North Maharashtra University, India Ayesh, Aladdin, De Montfort University, UK Bahreyni, Behraad, University of Manitoba, Canada Baoxian, Ye, Zhengzhou University, China Barford, Lee, Agilent Laboratories, USA Barlingay, Ravindra, RF Arrays Systems, India Basu, Sukumar, Jadavpur University, India Beck, Stephen, University of Sheffield, UK Ben Bouzid, Sihem, Institut National de Recherche Scientifique, Tunisia Benachaiba, Chellali, Universitaire de Bechar, Algeria Binnie, T. David, Napier University, UK Bischoff, Gerlinde, Inst. Analytical Chemistry, Germany Bodas, Dhananjay, IMTEK, Germany Borges Carval, Nuno, Universidade de Aveiro, Portugal Bousbia-Salah, Mounir, University of Annaba, Algeria Bouvet, Marcel, CNRS – UPMC, France Brudzewski, Kazimierz, Warsaw University of Technology, Poland Cai, Chenxin, Nanjing Normal University, China Cai, Qingyun, Hunan University, China Campanella, Luigi, University La Sapienza, Italy Carvalho, Vitor, Minho University, Portugal Cecelja, Franjo, Brunel University, London, UK Cerda Belmonte, Judith, Imperial College London, UK Chakrabarty, Chandan Kumar, Universiti Tenaga Nasional, Malaysia Chakravorty, Dipankar, Association for the Cultivation of Science, India Changhai, Ru, Harbin Engineering University, China Chaudhari, Gajanan, Shri Shivaji Science College, India Chen, Jiming, Zhejiang University, China Chen, Rongshun, National Tsing Hua University, Taiwan Cheng, Kuo-Sheng, National Cheng Kung University, Taiwan Chiang, Jeffrey (Cheng-Ta), Industrial Technol. Research Institute, Taiwan Chiriac, Horia, National Institute of Research and Development, Romania Chowdhuri, Arijit, University of Delhi, India Chung, Wen-Yaw, Chung Yuan Christian University, Taiwan Corres, Jesus, Universidad Publica de Navarra, Spain Cortes, Camilo A., Universidad Nacional de Colombia, Colombia Courtois, Christian, Universite de Valenciennes, France Cusano, Andrea, University of Sannio, Italy D'Amico, Arnaldo, Università di Tor Vergata, Italy De Stefano, Luca, Institute for Microelectronics and Microsystem, Italy Deshmukh, Kiran, Shri Shivaji Mahavidyalaya, Barshi, India Dickert, Franz L., Vienna University, Austria Dieguez, Angel, University of Barcelona, Spain Dimitropoulos, Panos, University of Thessaly, Greece Ding Jian, Ning, Jiangsu University, China Djordjevich, Alexandar, City University of Hong Kong, Hong Kong

Donato, Nicola, University of Messina, Italy Donato, Patricio, Universidad de Mar del Plata, Argentina Dong, Feng, Tianjin University, China Drljaca, Predrag, Instersema Sensoric SA, Switzerland Dubey, Venketesh, Bournemouth University, UK Enderle, Stefan, University of Ulm and KTB Mechatronics GmbH, Germany Erdem, Gursan K. Arzum, Ege University, Turkey Erkmen, Aydan M., Middle East Technical University, Turkey Estelle, Patrice, Insa Rennes, France Estrada, Horacio, University of North Carolina, USA Faiz, Adil, INSA Lyon, France Fericean, Sorin, Balluff GmbH, Germany Fernandes, Joana M., University of Porto, Portugal Francioso, Luca, CNR-IMM Institute for Microelectronics and Microsystems, Italy Francis, Laurent, University Catholique de Louvain, Belgium Fu, Weiling, South-Western Hospital, Chongqing, China Gaura, Elena, Coventry University, UK Geng, Yanfeng, China University of Petroleum, China Gole, James, Georgia Institute of Technology, USA Gong, Hao, National University of Singapore, Singapore Gonzalez de la Rosa, Juan Jose, University of Cadiz, Spain Granel, Annette, Goteborg University, Sweden Graff, Mason, The University of Texas at Arlington, USA Guan, Shan, Eastman Kodak, USA Guillet, Bruno, University of Caen, France Guo, Zhen, New Jersey Institute of Technology, USA Gupta, Narendra Kumar, Napier University, UK Hadjiloucas, Sillas, The University of Reading, UK Hashsham, Syed, Michigan State University, USA Hernandez, Alvaro, University of Alcala, Spain Hernandez, Wilmar, Universidad Politecnica de Madrid, Spain Homentcovschi, Dorel, SUNY Binghamton, USA Horstman, Tom, U.S. Automation Group, LLC, USA Hsiai, Tzung (John), University of Southern California, USA Huang, Jeng-Sheng, Chung Yuan Christian University, Taiwan Huang, Star, National Tsing Hua University, Taiwan Huang, Wei, PSG Design Center, USA Hui, David, University of New Orleans, USA Jaffrezic-Renault, Nicole, Ecole Centrale de Lyon, France Jaime Calvo-Galleg, Jaime, Universidad de Salamanca, Spain James, Daniel, Griffith University, Australia Janting, Jakob, DELTA Danish Electronics, Denmark Jiang, Liudi, University of Southampton, UK Jiang, Wei, University of Virginia, USA Jiao, Zheng, Shanghai University, China John, Joachim, IMEC, Belgium Kalach, Andrew, Voronezh Institute of Ministry of Interior, Russia Kang, Moonho, Sunmoon University, Korea South Kaniusas, Eugenijus, Vienna University of Technology, Austria Katake, Anup, Texas A&M University, USA Kausel, Wilfried, University of Music, Vienna, Austria Kavasoglu, Nese, Mugla University, Turkey Ke, Cathy, Tyndall National Institute, Ireland Khan, Asif, Aligarh Muslim University, Aligarh, India Kim, Min Young, Kyungpook National University, Korea South Sandacci, Serghei, Sensor Technology Ltd., UK

Ko, Sang Choon, Electronics and Telecommunications Research Institute, Korea South Kockar, Hakan, Balikesir University, Turkey Kotulska, Malgorzata, Wroclaw University of Technology, Poland Kratz, Henrik, Uppsala University, Sweden Kumar, Arun, University of South Florida, USA Kumar, Subodh, National Physical Laboratory, India Kung, Chih-Hsien, Chang-Jung Christian University, Taiwan Lacnjevac, Caslav, University of Belgrade, Serbia Lay-Ekuakille, Aime, University of Lecce, Italy Lee, Jang Myung, Pusan National University, Korea South Lee, Jun Su, Amkor Technology, Inc. South Korea Lei, Hua, National Starch and Chemical Company, USA Li, Genxi, Nanjing University, China Li, Hui, Shanghai Jiaotong University, China Li, Xian-Fang, Central South University, China Liang, Yuanchang, University of Washington, USA Liawruangrath, Saisunee, Chiang Mai University, Thailand Liew, Kim Meow, City University of Hong Kong, Hong Kong Lin, Hermann, National Kaohsiung University, Taiwan Lin, Paul, Cleveland State University, USA Linderholm, Pontus, EPFL - Microsystems Laboratory, Switzerland Liu, Aihua, University of Oklahoma, USA Liu Changgeng, Louisiana State University, USA Liu, Cheng-Hsien, National Tsing Hua University, Taiwan Liu, Songqin, Southeast University, China Lodeiro, Carlos, Universidade NOVA de Lisboa, Portugal Lorenzo, Maria Encarnacio, Universidad Autonoma de Madrid, Spain Lukaszewicz, Jerzy Pawel, Nicholas Copernicus University, Poland Ma, Zhanfang, Northeast Normal University, China Majstorovic, Vidosav, University of Belgrade, Serbia Marquez, Alfredo, Centro de Investigacion en Materiales Avanzados, Mexico Matay, Ladislav, Slovak Academy of Sciences, Slovakia Mathur, Prafull, National Physical Laboratory, India Maurya, D.K., Institute of Materials Research and Engineering, Singapore Mekid, Samir, University of Manchester, UK Melnyk, Ivan, Photon Control Inc., Canada Mendes, Paulo, University of Minho, Portugal Mennell, Julie, Northumbria University, UK Mi, Bin, Boston Scientific Corporation, USA Minas, Graca, University of Minho, Portugal Moghavvemi, Mahmoud, University of Malaya, Malaysia Mohammadi, Mohammad-Reza, University of Cambridge, UK Molina Flores, Esteban, Benemérita Universidad Autónoma de Puebla, Mexico Moradi, Majid, University of Kerman, Iran Morello, Rosario, DIMET, University "Mediterranea" of Reggio Calabria, Italy Mounir, Ben Ali, University of Sousse, Tunisia Mulla, Imtiaz Sirajuddin, National Chemical Laboratory, Pune, India Neelamegam, Periasamy, Sastra Deemed University, India Neshkova, Milka, Bulgarian Academy of Sciences, Bulgaria Oberhammer, Joachim, Royal Institute of Technology, Sweden Ould Lahoucine, Cherif, University of Guelma, Algeria Pamidighanta, Sayanu, Bharat Electronics Limited (BEL), India Pan, Jisheng, Institute of Materials Research & Engineering, Singapore Park, Joon-Shik, Korea Electronics Technology Institute, Korea South Penza, Michele, ENEA C.R., Italy Pereira, Jose Miguel, Instituto Politecnico de Setebal, Portugal Petsev, Dimiter, University of New Mexico, USA Pogacnik, Lea, University of Ljubljana, Slovenia Post, Michael, National Research Council, Canada Prance, Robert, University of Sussex, UK Prasad, Ambika, Gulbarga University, India Prateepasen, Asa, Kingmoungut's University of Technology, Thailand Pullini, Daniele, Centro Ricerche FIAT, Italy Pumera, Martin, National Institute for Materials Science, Japan Radhakrishnan, S. National Chemical Laboratory, Pune, India Rajanna, K., Indian Institute of Science, India Ramadan, Qasem, Institute of Microelectronics, Singapore Rao, Basuthkar, Tata Inst. of Fundamental Research, India Raoof, Kosai, Joseph Fourier University of Grenoble, France Reig, Candid, University of Valencia, Spain Restivo, Maria Teresa, University of Porto, Portugal Robert, Michel, University Henri Poincare, France Rezazadeh, Ghader, Urmia University, Iran Royo, Santiago, Universitat Politecnica de Catalunya, Spain Rodriguez, Angel, Universidad Politecnica de Cataluna, Spain Rothberg, Steve, Loughborough University, UK Sadana, Ajit, University of Mississippi, USA Sadeghian Marnani, Hamed, TU Delft, The Netherlands

Sapozhnikova, Ksenia, D.I.Mendeleyev Institute for Metrology, Russia Saxena, Vibha, Bhbha Atomic Research Centre, Mumbai, India Schneider, John K., Ultra-Scan Corporation, USA Seif, Selemani, Alabama A & M University, USA Seifter, Achim, Los Alamos National Laboratory, USA Sengupta, Deepak, Advance Bio-Photonics, India Shankar, B. Baliga, General Monitors Transnational, USA Shearwood, Christopher, Nanyang Technological University, Singapore Shin, Kyuho, Samsung Advanced Institute of Technology, Korea Shmaliy, Yuriy, Kharkiv National University of Radio Electronics, Ukraine Silva Girao, Pedro, Technical University of Lisbon, Portugal Singh, V. R., National Physical Laboratory, India Slomovitz, Daniel, UTE, Uruguay Smith, Martin, Open University, UK Soleymanpour, Ahmad, Damghan Basic Science University, Iran Somani, Prakash R., Centre for Materials for Electronics Technol., India Srinivas, Talabattula, Indian Institute of Science, Bangalore, India Srivastava, Arvind K., Northwestern University, USA Stefan-van Staden, Raluca-Ioana, University of Pretoria, South Africa Sumriddetchka, Sarun, National Electronics and Computer Technology Center, Thailand Sun, Chengliang, Polytechnic University, Hong-Kong Sun, Dongming, Jilin University, China Sun, Junhua, Beijing University of Aeronautics and Astronautics, China Sun, Zhiqiang, Central South University, China Suri, C. Raman, Institute of Microbial Technology, India Sysoev, Victor, Saratov State Technical University, Russia Szewczyk, Roman, Industrial Research Institute for Automation and Measurement, Poland Tan, Ooi Kiang, Nanyang Technological University, Singapore, Tang, Dianping, Southwest University, China Tang, Jaw-Luen, National Chung Cheng University, Taiwan Teker, Kasif, Frostburg State University, USA Thumbavanam Pad, Kartik, Carnegie Mellon University, USA Tian, Gui Yun, University of Newcastle, UK Tsiantos, Vassilios, Technological Educational Institute of Kaval, Greece Tsigara, Anna, National Hellenic Research Foundation, Greece Twomey, Karen, University College Cork, Ireland Valente, Antonio, University, Vila Real, - U.T.A.D., Portugal Vaseashta, Ashok, Marshall University, USA Vazquez, Carmen, Carlos III University in Madrid, Spain Vieira, Manuela, Instituto Superior de Engenharia de Lisboa, Portugal Vigna, Benedetto, STMicroelectronics, Italy Vrba, Radimir, Brno University of Technology, Czech Republic Wandelt, Barbara, Technical University of Lodz, Poland Wang, Jiangping, Xi'an Shiyou University, China Wang, Kedong, Beihang University, China Wang, Liang, Advanced Micro Devices, USA Wang, Mi, University of Leeds, UK Wang, Shinn-Fwu, Ching Yun University, Taiwan Wang, Wei-Chih, University of Washington, USA Wang, Wensheng, University of Pennsylvania, USA Watson, Steven, Center for NanoSpace Technologies Inc., USA Weiping, Yan, Dalian University of Technology, China Wells, Stephen, Southern Company Services, USA Wolkenberg, Andrzej, Institute of Electron Technology, Poland Woods, R. Clive, Louisiana State University, USA Wu, DerHo, National Pingtung University of Science and Technology, Taiwan Wu, Zhaoyang, Hunan University, China Xiu Tao, Ge, Chuzhou University, China Xu, Lisheng, The Chinese University of Hong Kong, Hong Kong Xu, Tao, University of California, Irvine, USA Yang, Dongfang, National Research Council, Canada Yang, Wuqiang, The University of Manchester, UK Ymeti, Aurel, University of Twente, Netherland Yong Zhao, Northeastern University, China Yu, Haihu, Wuhan University of Technology, China Yuan, Yong, Massey University, New Zealand Yufera Garcia, Alberto, Seville University, Spain Zagnoni, Michele, University of Southampton, UK Zeni, Luigi, Second University of Naples, Italy Zhong, Haoxiang, Henan Normal University, China Zhang, Minglong, Shanghai University, China Zhang, Qintao, University of California at Berkeley, USA Zhang, Weiping, Shanghai Jiao Tong University, China Zhang, Wenming, Shanghai Jiao Tong University, China Zhou, Zhi-Gang, Tsinghua University, China Zorzano, Luis, Universidad de La Rioja, Spain Zourob, Mohammed, University of Cambridge, UK

Sensors & Transducers Journal (ISSN 1726-5479) is a peer review international journal published monthly online by International Frequency Sensor Association (IFSA). Available in electronic and CD-ROM. Copyright © 2009 by International Frequency Sensor Association. All rights reserved.

Sensors & Transducers Journal

Contents Volume 5 Special Issue March 2009

www.sensorsportal.com

ISSN 1726-5479

Research Articles Foreword Pierre Payeur and Emil Petriu ............................................................................................................

1

An Omnidirectional Stereoscopic System for Mobile Robot Navigation Rémi Boutteau, Xavier Savatier, Jean-Yves Ertaud, Bélahcène Mazari ...........................................

3

Movement in Collaborative Robotic Environments Based on the Fish Shoal Emergent Patterns Razvan Cioarga, Mihai V. Micea, Vladimir Cretu, Emil M. Petriu.......................................................

18

A Multiscale Calibration of a Photon Videomicroscope for Visual Servo Control: Application to MEMS Micromanipulation and Microassembly Brahim Tamadazte, Sounkalo Dembélé and Nadine Piat..................................................................

37

A Study on Dynamic Stiffening of a Rotating Beam with a Tip Mass Shengjian Bai, Pinhas Ben-Tzvi, Qingkun Zhou, Xinsheng Huang....................................................

53

Towards a Model and Specification for Visual Programming of Massively Distributed Embedded Systems Meng Wang, Varun Subramanian, Alex Doboli, Daniel Curiac, Dan Pescaru and Codruta Istin ......

69

Feature Space Dimensionality Reduction for Real-Time Vision-Based Food Inspection Mai Moussa CHETIMA and Pierre PAYEUR .....................................................................................

86

Design and Analysis of a Fast Steering Mirror for Precision Laser Beams Steering Qingkun Zhou, Pinhas Ben-Tzvi and Dapeng Fan.............................................................................

104

Neural Gas and Growing Neural Gas Networks for Selective 3D Sensing: a Comparative Study Ana-Maria Cretu, Pierre Payeur and Emil M. Petriu...........................................................................

119

Authors are encouraged to submit article in MS Word (doc) and Acrobat (pdf) formats by e-mail: [email protected] Please visit journal’s webpage with preparation instructions: http://www.sensorsportal.com/HTML/DIGEST/Submition.htm International Frequency Sensor Association (IFSA).

Sensors & Transducers Journal, Vol. 5, Special Issue, March 2009, pp. 69-85

Sensors & Transducers ISSN 1726-5479 © 2009 by IFSA http://www.sensorsportal.com

Towards a Model and Specification for Visual Programming of Massively Distributed Embedded Systems Meng WANG, Varun SUBRAMANIAN and Alex DOBOLI Department of Electrical and Computer Engineering, State University of New York at Stony Brook, Stony Brook, NY 11794-2350, USA Tel.: +1-631-632-1611, fax: +1-631-632-8494 E-mail: [email protected]

Daniel CURIAC, Dan PESCARU and Codruta ISTIN Faculty of Automatics and Computers, “Politehnica” University Timisoara, Timisoara, 300223, Romania Tel.: +40-256-403227, fax: +40-256-403214 E-mail: [email protected] Received: 30 January 2009 /Accepted: 24 February 2009 /Published: 23 March 2009

Abstract: Massively distributed embedded systems are rapidly emerging as a key concept for many modern applications. However, providing efficient and scalable decision making capabilities to such systems is currently a significant challenge. This paper proposes a model and a specification language to allow automated synthesis of distributed controllers, which implement and interact through formalisms of different semantics. The paper refers to a case study to illustrate the main capabilities of the proposed concept. Copyright © 2009 IFSA. Keywords: Cyber-physical systems, Networks of sensing systems, Goal-oriented programming

1. Introduction Massively distributed embedded systems are rapidly emerging as a key enabling concept for many modern applications in infrastructure management, environmental monitoring, energy conservation, healthcare, homeland security, manufacturing, and many other [1-6]. This is due not only to sensing and electronic devices becoming extremely cheap and small in size (thus, deployable in large numbers) but also to the potential of having significantly superior decision making quality if related problems 69

Sensors & Transducers Journal, Vol. 5, Special Issue, March 2009, pp. 69-85

are tackled together instead of separately. Moreover, this helps improving the robustness of the systems as experience has shown that many disasters occur due to unaccounted correlations between the sub-systems. Providing reliable and efficient decision making capabilities to massively distributed embedded systems currently represents a main challenge [7]. The envisioned decision making paradigm is different from existing approaches, which either focus on centralized control or on local control [7], many times using ideas inspired from social or biological systems. Centralized control is well understood and reliable but does not scale well for large systems. In contrast, local control works well for large systems but with the exception of simple situations, its overall performance is hard to be captured. We argue for a distributed decision making mechanism in which parameterized control procedures implementing different strategies are dynamically introduced or removed depending on the specific optimization goals and operation conditions of the application. The controllers operate according to different models depending on the nature of their decision making. For example, reactive controllers use physical inputs for local control but operate under the constraints set by more abstract decision making procedures, which analyze broader situations based on aggregated (abstracted) data and procedures. The proposed decision making mechanism is flexible, scalable, and predictable. A key component of the approach is a suitable model and specification notation for massively parallel applications. This paper presents a novel control model and the related specification for developing massively distributed embedded applications. We believe that a main challenge in providing scalable descriptions for such applications is due to the wide variety of interactions that emerge among the composing subsystems, some of which are hard to anticipate a-priori, or might change their importance dynamically during execution. The proposed solution is to describe the nature of interactions that might occur among modules while leaving the task of optimally implementing these interactions to the compiler and execution environment. More specifically, the proposed model defines the operation goals of each sub-system (e.g., the criteria to be maximized or minimized during execution) and the physical capabilities of a module to achieve a certain goal (such as its maximum processing speed, highest bandwidth, etc). Different interaction types are introduced depending on the way the sub-systems influence each other’s goals and capabilities. The specification is compiled into a network of Decision Modules (DMs), which use reactive models to control decisions at the physical level, and more abstract formalisms, such as Task Graphs and Markov Decision Processes, to perform broader, more strategic decisions. The multisemantic DMs interact through (i) constraints by which the strategic modules restrict the reactive DMs to guarantee satisfaction of the global goals and (ii) feedback offered by the reactive DMs about the feasibility of the constraints. Section 2 summarizes the related work. Section 3 defines the distributed control model and Section 4 introduces the main language constructs. Section 5 illustrates the design of an application. Finally, conclusions are offered.

2. Related Work Developing concurrent processing systems is known to be a difficult and error-prone activity [8, 9, 10]. Existing programming models for concurrent processing differ depending on the granularity of the supported parallelism and communication [11]. Traditionally, concurrent threads synchronize and communicate through low level mechanisms, like semaphores, critical regions, send-receive primitives, etc. Interrupts are the main hardware support for communication [12]. More recent work suggests coordination models, in which interfaces define the proper cooperation and communication between threads [11-13]. Coordination models describe a common tuple space for placing and 70

Sensors & Transducers Journal, Vol. 5, Special Issue, March 2009, pp. 69-85

retrieving data (e.g., Linda [14]), adaptive parallelism by need-based activation and deactivation of threads (like Piranha [15]), manipulation of the tuple space including aggregation (i.e. Bonita [16]), constrained based communication (i.e., Law-Governed Linda [17]), moving objects from on space to another (e.g., Objective Linda [18]), and dynamic transformation of the tuple space objects (like in Gamma [19]). Compositional models, a special class of coordination models, include constructs for integrating components into bigger programs, such that the properties of the components are preserved. The integration constructs include logic clauses, like parallel AND operations in Strand [20], and higher-order functions [21, 22]. Interface-based design [23-26] is based on black-box models defining the system properties of the component interfaces, like arrival rate, latency, and capacity of the shared resources. The concept of Visual Programming (VP) was arguably proposed in the 80s [27], however, it is only recently that its advantages for embedded applications became apparent. VP languages have been proposed for applications like managing smart oilfields [4, 28], vehicle tracking [29, 30], contour finding [29], environmental monitoring [2], etc. Region Streams [29] is a functional macroprogramming language for sensor networks. Specification is based on successive filtering and functional processing of data pools. The data model is based on continuous data streams sampled from the environment and groups of nodes defined by their specific interests in space and over time. Language constructs enable aggregation of the data streams from a region and application of a function to the streams in a region. Abstract Task Graphs [2] is also a functional specification in which tasks sample from and place data into pools. There is no other interaction type between tasks. Channels are filters for associating only specific data from a pool to a task. Tasks are executed periodically or when input data is available. Semantic Streams [5] implements a query-based programming paradigm that fits well applications in which sensor networks operate as large distributed databases. Queries formulated as logic programs are converted by the compiler into a service graph for the network. Data sensing is modeled as streams. Other constructs include filtering by specifying properties of the streams, defining regions and sub-regions of the physical space, and performance requirements (e.g., quality of service). Kairos [30] proposes a set of language-independent extensions for describing global behavior of sensor networks controlled centrally. The extensions assume shared memory to allow any node to iterate through its neighbors and address arbitrary nodes. Similar to [5, 29], the data model of the proposed specification language is based on data pools and continuous data streams to the modules. However, it differs in that it focuses on optimal decision making in massively distributed environment and not on algorithmic descriptions. Therefore, we argue that the language is orthogonal to the existing notations as it concentrates on the interactions between groups of nodes, or nodes and environment, and less on the behavior of the individual nodes.

3. Goal-Oriented Model Cyber-physical Systems (CPS) are defined by a large set of interacting sub-systems. Each sub-system has well defined functionality and performance requirements, however, the interactions are dynamic, and their nature and intensity changes during execution. Fig. 1 illustrates a simplified CPS for intelligent traffic management. The goal of the system is to optimize the traffic flow of a region by adjusting the traffic-light characteristics to the specifics of the traffic flow. The example comprises of four sub-systems, the sub-system represented by the moving cars, the two data collection sub-systems (based on video cameras and sound-based tracking), and the sub-system formed by the traffic-light of the region. Each sub-system has a well defined functionality. • The image-based tracking system collects video images of traffic to get information like car position and speed, distance between cars, average number of cars passing through a region, size of car clusters, and so on. Video images have a certain precision in time and coverage: they are 71

Sensors & Transducers Journal, Vol. 5, Special Issue, March 2009, pp. 69-85

acquired at certain time intervals, and they might offer only a partial field-of-view covering. The sub-system can also detect events, like accidents or certain “special” cars, like police cars.

Fig. 1. Interactions in Cyber-Physical Systems.

• The sound-based tracking system follows the moving of specific cars, such as police cars, fire trucks, and ambulances. Having a tracking alternative can be very useful if the field of view to a vehicle is lost, or if the video-based tracking is too slow to meet the necessary timing constraints of the application. The system computes the position of a tracked car with a given precision and at certain intervals of time. The system also computes traffic statistics, like average speed, and detects traffic events, such as stopped cars. • The moving cars form a group of mobile embedded nodes. The mobile nodes can interact with each other both directly through a wireless, ad-hoc network, and indirectly by sharing the same physical space at a certain time. The node dynamics is described by attributes like speed and distance to neighboring cars. • The traffic-lights subsystem includes all traffic-lights of a region. The traffic-lights controllers are connected in a network to better coordinate their parameters offset, cycle and split times.

3.1. Sub-system Modeling: Functionality and Performance Each CPS sub-system is defined as shown in Fig. 2. The figure illustrates the main characteristics of the proposed distributed control concept and the related specification language. The wide geographical areas from which data is sampled define pools of heterogeneous data (e.g., images, temperature, humidity, number of moving vehicles, etc.). The association of data acquisition to the physical area is defined graphically, as shown in the left figure by the larger and smaller rectangles. For each defined area, the user specifies the goals that must be achieved by the distributed application, such as maximizing the traffic flow through the area, or keeping the pollution level below a preset limit.

72

Sensors & Transducers Journal, Vol. 5, Special Issue, March 2009, pp. 69-85

Fig. 2. Sub-system description for distributed decision making and control.

The execution platform is a massively distributed network of embedded controllers. Each controller node comprises hardware for sensing and actuation as well as processing, storage, and communication. In addition, hardware is reconfigurable, so that the available pool of resources at a node can be configured to meet different performance trade-offs, like variable processing speed, energy consumption, resolution, storage space, and so on. Hence, the model assumes that the functionality (algorithm) of each node is well defined and fixed but the node’s performance is parametric and dynamically changing at execution. The following three aspects are the core of the proposed model: • Separation of algorithmic aspects from goals: Algorithms describe the ontology of the application, and, over time, remain the same for all nodes in the network. In contrast, goals define optimization criteria, performance, safety, etc., depend on the specific execution platform and conditions, and change dynamically in time. While building algorithms is arguably more efficiently done by humans, finding the parameters for optimal execution is cumbersome but can be automated. • Description of interdependent, heterogeneous sub-systems: Global goals in large applications are likely to transcend different sub-system types. Also, the significance of the various related components can change over time. This invalidates static interaction schemes between sub-systems. • Avoiding explicit descriptions of synchronizations and data transfers: Explicit specification of internode communication reduces scalability and reusability. It is hard and unreliable to attempt capturing all possible interactions between the components of massively large scale applications. Instead, the design environment ought to identify the best interaction schemes between components, so that the overall goals as well as the goals of the modules are met. The proposed goal-oriented model comprises of separate Decision Modules (DMs) that operate to optimize a well-defined set of goals while the overall goals of the application are also being optimized. Each module executes a set of parameterized behaviors (algorithms) for which the parameters are automatically computed based on the information provided through the goal-oriented descriptions. DMs interact with each other only in small numbers but can perform global decisions by using data 73

Sensors & Transducers Journal, Vol. 5, Special Issue, March 2009, pp. 69-85

aggregated from large regions and by strategic decision methods (e.g., based on Markov Decision Processes). Fig. 3 illustrates a network, which comprises different decision making models (Finite State Machines, Markov Decision Processes, etc.). Similar to other specification languages for sensor networks, the proposed data model is based on data pools associated to regions and groups. Modules sample inputs from and generate outputs to a data pool for region. Regions represent continuous collections of tokens, such as a geographical area. Groups are discrete collections of tokens. Regions and groups can be associated to a specific physical area of the environment, or can be described by their defining properties. Efficient and robust decision making must be guaranteed for hard-to-predict conditions. The overall behavior should be capable of autonomously meeting performance requirements, e.g. real-time constraints, bandwidth limitations, energy constraints, speed requirements, precision, etc. The underlying decision making model (DMM) is based on a multi-semantic decision making networks that represents the overall application performance at different levels of abstraction and using different evaluation formalisms. The low levels employ reactive models, e.g., Finite State Machines, which are capable of tackling unexpected situations. The upper levels use less flexible models but with more predictable performance, such as Data Flow Graphs and Task Graphs. Thus, the semantic hierarchy offers a smooth transition from the fully reactive behavior at the embedded node level and the deterministic operation at the application level. The number of abstraction levels and the decision making models for each level depend on the application. The interaction between the different semantic models is achieved through top-down and bottom-up constraint transformation along the entire multi-semantic network: the top-down mechanism constrains the lower decision making levels by bounds introduced for their goals. As long as the low level operation stays within the bound it can be guaranteed that the overall performance is satisfied. The bottom-up mechanism signals when constraints (imposed by the upper levels) are cannot be satisfied by the lower levels. Fig. 3 shows a semantic hierarchy with three models, the bottom layers represent reactive behavior, and the top layers offer a performance predictive description of the system as Task Graphs (TGs) with data dependencies and Markov Decision Processes (MDPs).

Fig. 3. Specification and distributed control concept. 74

Sensors & Transducers Journal, Vol. 5, Special Issue, March 2009, pp. 69-85

The scheduling results of TGs are used to compute timing constraints for the reactive behavior of the bottom level DMs. As long as their reactive operation stays within these constraints, the overall timing requirement can be guaranteed. Bottom-up constraints express the amount of performance “violation” occurred at the physical level, and ought to be considered when re-computing the decision at the upper levels. The propagation of top-down and bottom-up constraints is performed continuously at runtime. The networked decision making in Fig. 3 operates as follows. The reactive DMs employ input sampling and event driven decision making mechanisms, such as Finite State Machines (FSMs). The reactive behavior switches from one task to another depending on the occurrence of events. For example, the system switches from S1 to S2, if event e1 occurs. Each embedded unit executes its own controller. Hardware reconfiguration offers the capability of selecting online task implementations with different parameters, like speed, energy, memory, communication bandwidth, etc. Several individual controllers might decide to collaborate by building collectively a shared description that defines how the individual controllers use jointly any shared resources to achieve common objectives, such as the access over time (schedule) of vehicles accessing intersections. For example, the shared description can be a Conditional Task Graph (CTG) [31], which includes decisions specific to different operation conditions, such as various traffic loads. Note that CTG decision making is at the level of small local areas rather than individual controllers. Next, the CTG descriptions of the correlated areas are used to build description covering broader areas, such as those at the Markov Decision Process level. This step distributes the overall goals into goals for the individual sub-systems using information from the Markov Decision Process level. The execution environment determines the structure of the distributed multi-semantic decision network as well as the node’s performance parameters. The network expands and shrinks depending on the nature of the application. Section 5 presents a case study for different decision making procedures.

3.2. Emergent Interactions The sub-systems interact in complex ways. Interactions include not only exchanging data between the sub-systems, like in traditional concurrent tasks, but also modifying the goals, constraints, and other parameters of the participating sub-systems. For example, the travel speed of cars sets the timing constraint of the image-based tracking sub-system, so that certain field of view coverage is guaranteed. The traffic characteristics (e.g., average speed and average number of passing cars) determine the timing parameters of the traffic-lights, like cycle time, offset and split time. The timing deadlines for “special” cars, like police cars and fire trucks, act as hard timing constraints for both the image-based and sound-based tracking sub-systems. Many interactions emerge for specific conditions. For example a high traffic load triggers the need to optimize the timing parameters of the traffic-lights, however, if the traffic load is very low then the default, periodic timing is sufficient, and hence there is no need of establishing interactions between the traffic-light and image-based tracking sub-systems. Also, the sound-based tracking sub-system needs to interact with the other sub-systems only in special conditions, such as if there is no direct sight of view or if the timing requirements are tight. In general, two sub-systems interact with each other whenever the activities of one have a significant impact on the behavior of the other sub-system. The significance of the impact is dynamic as it changes in time and space. Moreover, the identity of the interacting sub-systems is hard to predict a-priori due to the large number of potential interactions and the likelihood of adding new sub-systems to an existing CPS application. Incremental changes to the CPS functionality should be performed without recompiling the entire application, and without stopping the execution of the existing sub-systems. Adopting a conservative approach in implementing emergent interactions can lead to excessive communication overhead due to excessive synchronization and data exchange between sub-systems. 75

Sensors & Transducers Journal, Vol. 5, Special Issue, March 2009, pp. 69-85

We suggest the concept of interaction space as the main mechanism for defining and implementing emergent interactions between sub-systems. Interaction spaces are inspired from energy fields: interaction spaces define procedures for (i) adding and removing energy from a field, (ii) propagating energy, and (iii) measuring the energy without changing the status of the space. In addition, global energy sources define overall interaction needs for the sub-systems. The routines of each CPS subsystem have defined a “footprint” in terms of their impact on the interaction space. The footprint can change dynamically. Emergent interactions between sub-systems are identified by following the footprint of the sub-systems, such as a sub-system that generates significant amounts of energy interacts with another sub-system that is sensitive to the energy. Once an interaction is identified, dedicated communication channels are set-up between the sub-systems to provide the needed synchronization and data exchange with a reduced communication overhead. Fig. 1 illustrates the concepts. The space has multiple orthogonal directions, which may or may not correspond to the Cartesian directions. The directionality of interactions results from the way energy is added (input) and removed (output) from the space. The influence of each action ripples through the space along each direction depending on a propagation resistance along that direction. The influences propagate more if the resistance is lower. The propagation resistance can change dynamically depending on the impact of the sub-system that adds or removes energy. Also, the influence can be eliminated in case a sub-system must have only a local influence. A possible representation of an interaction space can be based on electrical fields. For our example, the cars entering the region can be interpreted as being analogous to electrical charge, thus adding energy to the electrical field corresponding to the interaction space. The cars leaving the region represent removing of energy from the field. The two tracking sub-systems measure the energy of the interaction space. Each route of the region defines a direction of the space. The traffic characteristics along a route, e.g., the average speed of cars, define the propagation resistance along that direction. The propagation resistance changes dynamically for each car or group of cars depending on the moving characteristics of the cars. Each traffic-light behaves as a switch that controls the moving of the charge particles along the existing routes, hence control the propagation resistance along directions. This representation can be used to identify emergent interactions between separate sub-systems and also between the components of the same sub-system. For example, if the traffic flow increases significantly along a certain route then the related energy increase affects a certain area of the region. The traffic-lights in that area must be coordinated with each other as they control the propagation resistance. Another example refers to the emerging interactions between video-based and sound-based tracking system. The coverage dropping of the video-based sub-system is described as a steep increase in the propagation resistance along the route. The resistance increase can be compensated through a similar decrease if the sound-based tracking sub-system can provide the necessary tracking. Hence, an emerging interaction is established between the two sub-systems. Note that the interaction space representation defines only a qualitative view of a CPS application, hence it is not used for quantitative reasoning, such as for detailed decision making and control at the level of the individual sub-systems.

4. Main Specification Constructs The main constructs of the proposed goal-oriented specification notation are illustrated in Fig. 4. The basic specification entity is a Decision Module (DM), which corresponds to local points, physical areas, zones, or larger regions. As explained in Section 3, each DM executes specific applicationspecific algorithms for sensing, processing, and actuation. This part is internal to each DM and is specified using an existing programming language, like C++ or Java. The focus of the proposed specification notation is on describing the local and global goals and the nature of interactions between modules. The compiler automatically produces the actual interaction mechanism, such communication mechanism and parameters. For simplicity, the presentation is descriptive instead of a formal one. 76

Sensors & Transducers Journal, Vol. 5, Special Issue, March 2009, pp. 69-85

Fig. 4. Specification for scalable decision making.

Fig. 4 illustrates that a DM has four main parts: inputs, outputs, capabilities, and goals: • Inputs: DM inputs either sample the pool of data associated to a physical region or are interaction data coming from other DMs. The acquisition semantics of an input can be continuous time, discrete time, or event driven. Moreover, inputs can refer only to certain facets of physical signals, e.g., voltage, current, phase, and frequency of a signal. The user can also specify aggregation of signals over time and space (integrals for continuous inputs and sums for discrete signals) and rate of change (sensitivities) with respect to time, space, or other signals (derivates for continuous inputs and differences for discrete inputs). • Outputs: DM outputs relate to the outputs produced by the algorithm of a module. Outputs include physical actuation and control signals. • Capabilities: Each module has a limited set of physical capabilities, such as the highest amount of service requests it can service, local memory, processing speed, energy resources, communication distance and bandwidth, and so on. These capabilities affect the ability of a node to maximize its local goals, and also percolate in influencing the quality of the optimal value that is reached for the overall application. For example, the low processing speed of a node or lack of memory might restrict the amount of alternatives that are locally analyzed, and affect the quality of the optimization process conducted by the node. The user can define attributes, like the upper and lower bounds and average value in time and space of a capability. • Goals: The part indicates the goals to be optimized by each DM. Goals can be expressed either as maximizing or minimizing a cost function, or as a constraint satisfaction requirement, in which the goal expression must either exceed or fall below a threshold value. The cost function is defined over outputs and capabilities. Components interact through interactions. The interactions are identified dynamically based on the interaction space concept discussed in Subsection 3.2. The following kinds of interactions can be established between DMs in the proposed notation: • Collaborative interactions: Collaborative interactions are set up between DMs, which have nonconflicting (non-competing) goals. Modules interact with each other through inputs and outputs, 77

Sensors & Transducers Journal, Vol. 5, Special Issue, March 2009, pp. 69-85

e.g., one module produces outputs to the pools that serve as inputs to another module. The goals and capabilities of the modules are not affected by this kind of interactions. • Competing interactions: Such interactions are between DMs with competing goals, like optimizing the goal of one DM affects adversely the optimization of the other DM. Similarly to collaborative interactions, modules interact through their inputs and outputs but cannot change their goals or capabilities. • Guiding interactions: Guided interactions are between DMs at consecutively higher levels in the semantic hierarchy. DMs at upper levels generate outputs that are used to set the goals of the modules at lower levels. This way the first modules are steering the goals and hence the behavior of the latter modules. • Enabling interactions: Enabling interactions are information transfers from lower (reactive) semantic levels to the upper levels. A lower DM transmits information about its capabilities to an upper DM, so that the second can use this knowledge during a decision making that might affect the goals set for the lower DM through guided interactions. Enabling interactions are the information links through which upper DMs acquire knowledge about the actions that are ongoing in the real world. In the proposed specification notation, the nature of interactions is dynamic and does not explicitly state the DMs participating to it. Instead, the user describes the conditions (thresholds) under which DMs start to interact with each other, like the change of an input, exceeding a threshold value, exceeding certain capabilities, etc. Interactions are described using Interaction Modules (IMs). IMs define any data transformation (such as data aggregation, filtering, etc.) that is needed if DMs of different formalisms are interacting. Similar to DMs, IMs have goals and capabilities. Moreover, the user must specify for each of the four interaction types, the formal mechanism (TGs, MDPs, etc.) used to resolve the interaction.

5. A Case Study and Insight into Specification Compiling This section offers insight onto how the proposed model is used for efficient decision making and implementation. The discussion is based on a case study on coordinated traffic management in urban areas. The section discusses decision making strategies at different levels of abstraction, such as the traffic-light level, zone level, and area level. Different models and methods are used in each case. The section also highlights the interactions between these levels. The traffic-light management sub-system has the goal to maximize the total number of vehicles passing through the managed area while the delay of a car is kept below the preset limit T. This constraint guarantees that cars on the less traveled directions are not indefinitely delayed by the vehicles moving on busier directions. Fig. 5a presents a simple street network. Arcs indicate the traffic flow. Traffic-lights (shown as small, black bubbles in the figure) are positioned at each intersection. The traffic-light parameters, e.g., cycle time, split time and offset, must be controlled by the subsystem to maximize the traffic flow. The traffic management sub-system has a hierarchical structure so that decision making scales well over large geographical areas. The lower structures use semantic models that are based on physical parameters, such as the vehicle speed and inter-vehicle distance. The upper levels employ semantic models that provide an aggregated, more strategic view of the traffic flow. The hierarchy is as follows: • Each traffic-light is controlled by a local DM called Intersection Traffic Unit (ITU). ITUs have reactive behavior based on FSMs with parameters optimized for the maximum traffic flow. • Neighboring ITUs form Zone Traffic Coordination Units (ZTCUs) that coordinate the ITUs. Using information from the individual ITUs, ZTCUs compute the ITU parameters that offer the best 78

Sensors & Transducers Journal, Vol. 5, Special Issue, March 2009, pp. 69-85

overall traffic flow for the set of coordinated ITUs. The ITU interactions are emergent: ITUs interact with each other only if cars are passing through the intersection. • The ZTCUs of an area are coordinated by the higher-level DM called Area Traffic Control Unit (ATCU). ATCUs formulate a global, aggregated view about the traffic flow, and calculate the ZTCU parameters that optimize the traffic flow of the entire area. Fig. 5 and Fig. 6 illustrate the first two lower levels of the multi-semantic, hierarchical decision making model: the FSM of the ITU and the Conditional Task Graph (CTG) of the ZTCU. The figures also show the interaction between the two levels. Each ITU implements a FSM with three states, Red, Yellow, and Green. DMs are executed by each embedded node. Assuming cyclic traffic signals, the arcs in Fig. 6b define the state transitions, which are taken after the FSM spends Ti,j time units in the current state (the time is called split time). The time intervals are specific to each ITU and transition, and are fixed by the ZTCU to optimize the traffic flow through the zone. The schedule of the traffic signal repeats after a period called cycle time. The neighboring ITUs interact with each other to decide the delay between the monitored intersections (the delay is called offset). The decision making methods of the ITUs and ZTCUs are discussed next.

Fig. 5. ITU-level behavior: a) simple street network, b) time behavior of ITU and c) traffic parameter estimation.

Fig. 6. FSM – CTG interaction: a) CTG for traffic in Figure 5(a) and b) ZTCU-level CTG scheduling table and interaction with ITU-level FSM.

79

Sensors & Transducers Journal, Vol. 5, Special Issue, March 2009, pp. 69-85

Fig. 5 illustrates the physical-level behavior of the traffic-light sub-system. The figure indicates the sequence of states along the directions In1 and In3 in Fig. 5a. Fig. 5b shows the time behavior of the ITUs (for each traffic-signal). N1 vehicles are passing along direction In1 as long as the traffic-light is in state Green1, and N2 vehicles are passing along direction In3 as long as the traffic light is in state Green2. The discussion assumes that no cars are passing if the current state is either Redi or Yellowi. The buffer states Redc,i and Red*c,i were introduced to guarantee that no vehicles are moving through the intersection during the switching to the states Greeni along any of the two directions. The moving of the vehicles is characterized by two main parameters, the average time Texi required to traverse the route portion I, and the number of vehicles Nout,j that move along direction j through an intersection. Using Fig. 5c, we estimate the number of vehicles that are moving along Direction In1 and are going through the following intersection during the next state Green of the traffic-light. The same reasoning can be applied for all other directions. Value Nw is the number of vehicles waiting in the queue of the next traffic-light. Value Nm is the number of cars arriving at the intersection along the observed direction. Time Ttravel is the time that cars need to move between the two consecutive trafficlights. Time ∆ is the delay between the traffic-lights. There are two scenarios: (i) if the Nm cars reach the next intersection before the Nw cars pass through, and (ii) if the cars reach after the waiting queue was cleared. In the first case, some of the Nm cars experience delays, while in the second case the in-coming cars are not delayed. The two cases are analyzed next. The time needed for all Nw vehicles to pass through the intersection is equal to Tthrough = Nw/nL + ∆, where nL is the number of vehicles passing per unit of time after vehicles were stopped. The first case corresponds to the conditions Ttravel ≥ Tthrough and Tthrough ≤ TGreen. Otherwise, the second case is performed. The two cases are analyzed next. Case (i): The value Nleft = Nw - nL (Ttravel - ∆) is the number of cars that are still in the queue when the pool of Nm cars reaches the intersection. Two situations can be identified. First, if nIn > nL, where nIn is the number of cars arriving along the observed direction in a unit of time then all Nm cars are delayed. Hence, the number of cars passing through the intersection is Nout = min (Nw + Nm, nL TGreen). The number of cars remaining in the waiting queue, hence representing the value Nnextw for the next trafficlight period, is equal to Nnextw = max (0, Nw + Nm - nL TGreen). Second, if nIn < nL then only some of the Nm cars are delayed in the waiting queue of the traffic-light, while the following cars pass through without delay before the traffic-light changes states. The time requited to clear the waiting queue is equal to Tclear = Nw/(nL - nIn). If Tclear ≥ TGreen then only a fraction of the waiting cars can pass through before the traffic-light changes states. Hence, Nout = min (Nw nL/(nL - nIn), nL x TGreen) and Nnextw = Nw + Nm - Nout. If Tclear < TGreen then the number of passing-through cars is equal to Nout = Nw nL/(nL - nIn) + min (nH (TGreen - Tclear), Nm). Value nH is the number of cars passing through per unit of time, if the cars are not placed in the waiting queue at the intersection. The cars remaining in the waiting queue for the next period are Nnextw = Nw + Nm - Nout. Case (ii): The time required to move the Nw vehicles through the intersection is equal to Tclear = Nw/ nL. If Tclear ≥ TGreen then only a fraction of the Nw vehicles are moving through, hence, Nout = nL TGreen. The number of cars that are placed in the waiting queue for the next traffic-light period is equal to Nnextw = Nw - nL TGreen + Nm. If Tclear < TGreen then Nout = Nw + min (Nm, nH (TGreen - Tdelay + ∆)), and Nnextw = Nw + Nm - Nout. The analysis shows that the ZTCU-level behavior experiences multiple modes of operation depending on the traffic conditions. To represent the conditional behavior, the ZTCU functionality is expressed as a Conditional Task Graph (CTG) [31], which describes the activities performed by the moving vehicles as they pass through the zones controlled by the ITUs. The activities are selected depending on conditions such as the traffic load. Hence, CTGs capture all situations in which a car can be involved. The analysis method assumes that one car enters the corresponding zone from each of its 80

Sensors & Transducers Journal, Vol. 5, Special Issue, March 2009, pp. 69-85

directions. For example, in Fig. 5a, one car enters along directions In1, In2, and In3. The related activities include passing through the intersection, moving to the next intersection, and so on. We assumed single-lane roads. This implies that two cars coming through the intersections ITU1 and ITU2 cannot move simultaneously. Fig. 6a presents the CTG for the zone in Fig. 5a. There is a parallel thread for each direction of the route. The graph presents the activities for moving a car through the zone. Tasks T1 and T2 describe the traveling of a vehicle entering the zone from the left, tasks T3 and T4 model the moving of a vehicle coming from the top-left road, tasks T6 and T7 represent a car moving from top right, and so on. Each of the activities is characterized by a typical execution time Texk that is estimated based on the data collected locally by the sensors at the traffic-lights, e.g., the video-based and sound-based tracking sub-systems. The execution times differ for different traffic scenarios, such as slow traffic (branch Li) and heavy traffic (branch Hi). The semantics of CTGs states that one and only one of the two branches Li and Hi originating at a conditional node is executed for each traversal of the graph. This semantics defines a conditional (multimode) behavior. Certain road sections (shown as dark areas in the figure) act as shared resources for two or more cars, while other sections are dedicated resources. In the graph, the tasks sharing the same resources are grouped in the two clusters marked with dashed and dotted lines. In addition, the execution time ∆ of the dummy task T10 defines the offset time between the two traffic-lights. Note that the parallel threads through the CTG interact through the number Nw of cars waiting before a traffic-light. These cars were “placed” in the waiting queue by the previous thread. Hence, finding fast schedules requires not only minimizing the total execution time of the current flow but also controlling the queue lengths. ZTCU computes the optimal scheduling of the tasks Ti to maximize the number of cars that move through the zone over a period of time, and/or minimize the overall time taken to cars to move through the zone. Different traffic conditions determine the values of the task attributes, like the execution time Tiex. The schedule lengths are used to compute the optimal cycle and split times of each ITU, and also find the delay values between successive traffic-lights. This step is discussed later in the section. Each schedule corresponds to traffic conditions described by the conditional behavior of the CTG. For example, if the traffic is slow on Segment 1 then the branch labeled L is taken after task T1 is completed, hence task T2 is performed. If the traffic load is high then branch H is selected, and task T2' is pursued. Tasks T2 and T2' can never be performed simultaneously as they correspond to mutually excluding conditions (the traffic intensity cannot be simultaneously low and high). This generates 16 different situations for the case in Fig. 5, such as schedule S(L,L,L) with the total time Ttot(L,L,L) for the situation in which the traffic load is light for all three intersections. The concrete task scheduling for CTGs can be computed with scheduling algorithms similar to the one in [31]. Fig. 6b presents the interaction between the FSMs of the ITUs at each traffic-light and the CTG scheduling at the ZTCU. The structure of the scheduling table computed online by the ZTCU is shown in the figure. The table includes the schedules for the scenarios that can emerge in traffic. The table in the figure presents two situations: one in which traffic is light along all directions, and the second in which the traffic along Direction 2 is high, and it remain low along the others. The boxes in gray indicate idle times along a direction. The information in the CTG scheduling table is mapped down onto the FSMs of each individual traffic-light. For example, for Scenario 1 (in which the traffic intensity is low along all directions), the time required for performing task T3 plus the following idle time generate the timing constraint ∆Tx, which states that after that amount of time the FSM must be in state Green, thus allowing traffic to progress along Direction 2. Similarly, for Scenario 2, the time for task T1 plus the following idle time define the constraint ∆Ty, which states that the traffic signal must be in state Red after time ∆Ty, which allows traffic to progress along Direction 1. As long as each traffic signal meets its local timing constraints (∆Tx and ∆Ty), the behavior at the ZTCU-level meets the traffic schedules specified in the ZTCU table. 81

Sensors & Transducers Journal, Vol. 5, Special Issue, March 2009, pp. 69-85

Fig. 7a introduces the more abstract decision making level, called Area Traffic Coordination Unit (ATCU). Its procedure is based on Markov Decision Processes (MDP). The figure also presents the connection of ATCU to the lower-level CTGs. For an estimated number of vehicles that pass through the area, each ZTCU calculates the optimal activity schedules for the predicted traffic scenario. If the timing delay TZTCUi of the ZTCU schedule i exceeds the timing constraint for traversing the area then the traffic-light controllers are adjusted so that the new traffic scenarios produce schedules that meet the constraint. Figure 7(b) offers a pictorial description of Continuous-Time Markov Chains (CTMCs):

CTMC = ( S , A, A(i ), p, K , r )

(1)

where S is the set of states, A is the action set, and A(i) are the actions associated to state i in S. p(i,j,a) is the transition rate if action a is chosen for transitioning from state i to state j. K is the number of reward criteria, and rk(i,a) is the reward rate if action a is selected for state i. xi is the steady-state probability of state i.

Fig. 7. CTG - MDP interaction: a) interaction between ZTCU-level CTG and ATCU-level MDP and b) MDP definition.

CTMC are employed at the ATCU-level to control the moving of vehicles over time so that the traffic load is maximized. The interaction between ZTCUs models the fact that the number Nout of outgoing vehicles for a ZTCU is equal to the number Nm of ingoing vehicles for the next ZTCU. Hence, for the situation in Figure 8 (a), Nout,1 = Nm,2 and Nout,2 = Nm,3. In addition, the behavior of each ZTCU is expressed as a separate CTMC, as presented in Fig. 8 b.

Fig. 8. MDP-level decision making: a) sequence of three ZTCUs and b) interaction of the related MDPs. 82

Sensors & Transducers Journal, Vol. 5, Special Issue, March 2009, pp. 69-85

The states of the CTMC correspond to the number of vehicles waiting in the k queues of the zone monitored by each ZTCU. The transition rate pi,j,Nm,k between any pair of states i and j depends on the number Nm,k of vehicles entering the zone. Hence, the values Nm,k act as actions for the CTMC. The transition rate is equal to the frequency of having Nm incoming vehicles. The reward criteria are the execution time required for the vehicles to reach the end of the zones (the execution time corresponds to the lengths of the schedules computed by the ZTCU) and the number Nout of vehicles that exit the zone. The following set of equations is formulated for each ZTCU I:

∑

∀Nm

p(i, i, Nm) xi,Nm - ∑ ∑ p(i, j, Nm) xi, Nm = 0, ∀j ∈ SI,

(2)

j∈SI ∀Nm

∑∑

xi,Nm = 1, ∀i ∈ SI, ∀ Nm

(3)

i∈SI Nm

xi,Nm ≥ 0, ∀i ∈ SI, ≥ 0, ∀ Nm

(4)

The set of equations must be solved so that the total number of vehicles output at the last ZTCU (ZTCU3 in Figure 8(a)) is maximized: max

∑∑

Nout (i, Nm) xi,Nm

(5)

i∈I Nm

The steady-state probabilities are computed by solving the set of equations. The probabilities are utilized to calculate the percentage of time that each alternative is used at the ZTCU-level.

6. Conclusions Massively distributed embedded systems are rapidly emerging as a breakthrough concept for many modern applications. However, providing efficient and scalable decision making capabilities to such systems is currently a significant challenge. The paper proposes a model and specification language to allow automated synthesis of distributed controllers, which implement and interact through models of different semantics. Scalability of descriptions is realized through defining the nature of interactions that can occur among decision modules while leaving to task of optimally implementing these interactions by the execution environment. The notation defines the operation goals of each sub-system (e.g., the criteria to be maximized or minimized during operation) and the physical capabilities of a module to achieve a certain goal. Different interaction types are introduced depending on the way subsystems influence each other their goals and capabilities. The paper refers to a case study to illustrate the model and the related decision making steps. Compared to similar work, the proposed model and specification notation differs in that they focus on optimal goal satisfaction in massively distributed systems and not on algorithmic descriptions. Therefore, the language is orthogonal to existing notations as it concentrates on the interactions between groups of nodes, or nodes and environment and less on the behavior of the individual nodes. This simplifies specification and helps scalable decision making.

83

Sensors & Transducers Journal, Vol. 5, Special Issue, March 2009, pp. 69-85

References [1]. C. Batini, E. Bertini, M. Comerio, A. Maurino, G. Santucci, Visual languages and quality evolution in multichannel adaptive information systems, Journal of Visual Languages & Computing, Vol. 18, Issue 5, 2007, pp. 513-522. [2]. A. Bakshi, V. Prasanna, J. Reich, D. Larner, The Abstract Task Graph: A methodology for architectureindependent programming of networked sensor systems, in Proceedings of End-to-End, Sense-and-respond Systems, Applications and Services, Seattle, Washington, 2005, pp. 19-24. [3]. L. Camara, E. Jungert, A visual query language for dynamic process applied to a scenario driven environment, Journal of Visual Languages & Computing, Vol. 18, Issue 3, 2007, pp. 315-338. [4]. R. Soma, A. Bakshi, V. Prasanna, A semantic framework for integrated asset management in smart oilfields, in Proceedings of the IEEE Symposium on Cluster Computing and Grid, Rio de Janeiro, Brazil, May 2007, pp. 119-126. [5]. K. Whitehouse, F. Zhao, J. Liu, Semantic streams: a framework for declarative queries and automatic data interpretation, Technical Report, Microsoft Research, MSR-TR-2005-45, 2005. [6]. A. Bakshi, V. Prasanna, A. Ledeczi, Milan: A model based integrated simulation framework for design of embedded systems, ACM Sigplan Notices, Snowbird, Utah, 2001, pp.82-93. [7]. K. Passino, Biomimicry for Optimization, Control, and Automation, Springer, 2005. [8]. E. Lee, The problem with threads, IEEE Computer, Vol. 39, Issue 5, 2006, pp. 33-42. [9]. H. Sutter, J. Larus, Software and the concurrency revolution, ACM Queue, Vol. 3, Issue 7, 2005, pp. 54-62. [10].N. Zeldovich, A. Yip, F. Dabek, R. Morris, D. Mazieres, F. Kaashoek, Multiprocessor support for eventdriven programs, in Proceedings of the USENIX Annual Technical Conference (USENIX’ 03), San Antonio Texas, June 2003. [11].G. Papadopoulos, F. Arbab, Coordination models and languages, Advances in Computers - The Engineering of Large Systems, Academic Press, Vol. 46, 1998, pp. 329-400. [12].E. Lee, Cyber physical systems: Design challenges, Technical Report No. UCB/EECS-2008-8, University of California at Berkeley, 2008. [13].G. Papadopoulos, A. Stavrou, O. Papapetrou, An implementation framework for software architectures based on the coordination paradigm, Science of Computer Programming, Vol. 60, Issue 1, 2006, pp. 27-67. [14].S. Ahuja, N. Carrieri, D. Gelernter, Linda and friends, IEEE Computer, Vol. 19, Issue 8, 1986, pp. 26-34. [15].D. Gelernter, D. Kaminsky, Supercomputing out of the recycled garbage: Preliminary experience with Piranha, in Proceedings of the ACM International Conference on Supercomputing, Washington DC, 1992, pp. 417-427. [16].A. Rowstron, A. Wood, Bonita: A set of tuple space primitives for distributed coordination, in Proceedings of the International Conference on Systems Sciences, Waikoloa, Hawaii, Vol. 1, 1997, pp. 379-388. [17].N. Minsky, J. Leichter, Law-governed Linda as a coordination model, in Object-Oriented Models and Languages for Concurrent Systems, LNCS 924, Springer Verlag, 1994, pp.125-145. [18].T. Kielmann, Designing a coordination model for open systems, in Proceedings of International Conference on Coordination Models, Languages and Applications, LNCS 1061, Springer Verlag, 1996, pp. 267-284. [19].J.-P. Banatre, D. Le Metayer, Gamma and the chemical reaction model: Ten years after, in Coordination Programming: Mechanisms, Models and Semantics, World Scientific, 1996, pp.1-39. [20].I. Foster, Compositional parallel programming languages, ACM Transactions on Programming Languages and Systems, Vol. 18, Issue 4, 1996, pp. 454-476. [21].M. Cole, Algorithmic Skeletons: Structured Management of Parallel Computation, MIT Press, 1989. [22].D. B. Skillicorn, Towards a higher level of abstraction in parallel programming, in Proceedings of Programming Models for Massively Parallel Computers, Berlin, Germany, 1995, pp. 78-85. [23].L. de Alfaro, T. Henzinger, Interface-based design, Engineering Theories of Software-Intensive Systems, NATO Science Series: Mathematics, Physics, and Chemistry, Vol. 195, Springer, 2005, pp. 83-104. [24].T. Henzinger, B. Horowitz, B. C. Kirsch, Giotto: A time-triggered language for embedded programming, Proceedings of IEEE, Vol. 91, Issue 1, 2003, pp. 84-99. [25].T. Henzinger, S. Matic, An interface algebra for real-time components, in Proceedings of the Annual RealTime and Embedded Technology and Applications Symposium, San Jose, California, 2006, pp. 253-266. [26].E. Wandeler, L. Thiele, Real-time interfaces for interface-based design of real-time systems with fixed priority scheduling, in Proceedings of International Conference on Embedded Software (EMSOFT), Jersey City, New Jersey, 2005, pp. 80-89. [27].W. Johnston, Advances in dataflow programming languages, ACM Computing Survey, Vol. 36, Issue 1, 84

Sensors & Transducers Journal, Vol. 5, Special Issue, March 2009, pp. 69-85

2004, pp. 1-34. [28].C. Zhang, A. Bakshi, V. Prasanna, ModelML: a markup language for automatic model synthesis, in Proceedings of the IEEE Conference o Information Reuse and Integration, Las Vegas, Nevada, 2007, pp. 317-322. [29].R. Newton, M. Welsh, Region Streams: Functional macroprogramming for sensor networks, in Proceedings of the International Workshop on Data Management for Sensor Networks, Toronto, Canada, 2004, pp. 78-87. [30].K. Gummadi, O. Gnawali, R. Govindan, Macro-programming wireless sensor networks using Kairos, in Proceedings of the International Conference on Distributed Computing in Sensor Systems, 2005, pp. 126-140. [31].P. Eles, A. Doboli, Z. Peng, P. Pop, Scheduling with buss access optimization for distributed embedded Systems, IEEE Transactions on VLSI Systems, Vol. 8, No. 5, 2000, pp.472-491.

___________________ 2009 Copyright ©, International Frequency Sensor Association (IFSA). All rights reserved. (http://www.sensorsportal.com)

85

Sensors & Transducers Journal

Guide for Contributors Aims and Scope Sensors & Transducers Journal (ISSN 1726-5479) provides an advanced forum for the science and technology of physical, chemical sensors and biosensors. It publishes state-of-the-art reviews, regular research and application specific papers, short notes, letters to Editor and sensors related books reviews as well as academic, practical and commercial information of interest to its readership. Because it is an open access, peer review international journal, papers rapidly published in Sensors & Transducers Journal will receive a very high publicity. The journal is published monthly as twelve issues per annual by International Frequency Association (IFSA). In additional, some special sponsored and conference issues published annually.

Topics Covered Contributions are invited on all aspects of research, development and application of the science and technology of sensors, transducers and sensor instrumentations. Topics include, but are not restricted to: • • • • • • • • • • • • •

Physical, chemical and biosensors; Digital, frequency, period, duty-cycle, time interval, PWM, pulse number output sensors and transducers; Theory, principles, effects, design, standardization and modeling; Smart sensors and systems; Sensor instrumentation; Virtual instruments; Sensors interfaces, buses and networks; Signal processing; Frequency (period, duty-cycle)-to-digital converters, ADC; Technologies and materials; Nanosensors; Microsystems; Applications.

Submission of papers Articles should be written in English. Authors are invited to submit by e-mail [email protected] 6-14 pages article (including abstract, illustrations (color or grayscale), photos and references) in both: MS Word (doc) and Acrobat (pdf) formats. Detailed preparation instructions, paper example and template of manuscript are available from the journal’s webpage: http://www.sensorsportal.com/HTML/DIGEST/Submition.htm Authors must follow the instructions strictly when submitting their manuscripts.

Advertising Information Advertising orders and enquires may be sent to [email protected] Please download also our media kit: http://www.sensorsportal.com/DOWNLOADS/Media_Kit_2008.pdf